def _init_training(self, das_file, ttree_file, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
"""
# read input
log_info('Reading DAs from ' + das_file + '...')
das = read_das(das_file)
log_info('Reading t-trees from ' + ttree_file + '...')
ttree_doc = read_ttrees(ttree_file)
trees = trees_from_doc(ttree_doc, self.language, self.selector)
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DialogueAct()
empty_da.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([self.tree_embs.get_embeddings(tree) for tree in self.train_trees])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.X = [self.tree_feats.get_features(tree, {}) for tree in self.train_trees]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [self.da_feats.get_features(None, {'da': da}) for da in self.train_das]
self.y = self.da_vect.fit_transform(self.y)
# initialize I/O shapes
self.input_shape = [list(self.X[0].shape)]
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
def _init_training(self, das_file, ttree_file, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
"""
# read input
log_info('Reading DAs from ' + das_file + '...')
das = read_das(das_file)
log_info('Reading t-trees from ' + ttree_file + '...')
ttree_doc = read_ttrees(ttree_file)
trees = trees_from_doc(ttree_doc, self.language, self.selector)
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([
self.tree_embs.get_embeddings(tree)
for tree in self.train_trees
])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False,
binarize_numeric=True)
self.X = [
self.tree_feats.get_features(tree, {})
for tree in self.train_trees
]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [
self.da_feats.get_features(None, {'da': da})
for da in self.train_das
]
self.y = self.da_vect.fit_transform(self.y)
# initialize I/O shapes
self.input_shape = [list(self.X[0].shape)]
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
def __init__(self, cfg):
self.cfg = cfg
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.mode = cfg.get('mode', 'tokens' if cfg.get('use_tokens') else 'trees')
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.cur_da = None
self.cur_da_bin = None
self.delex_slots = cfg.get('delex_slots', None)
if self.delex_slots:
self.delex_slots = set(self.delex_slots.split(','))
def __init__(self, cfg):
super(EmbNNRanker, self).__init__(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.normgrad = cfg.get('normgrad', False)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
# 'emb' = embeddings for both, 'emb_trees' = embeddings for tree only, 1-hot DA
# 'emb_tree', 'emb_prev' = tree-only embeddings
self.da_embs = cfg.get('nn', 'emb') == 'emb'
self.tree_embs = TreeEmbeddingExtract(cfg)
if self.da_embs:
self.da_embs = DAEmbeddingExtract(cfg)
else:
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.vectorizer = None
def __init__(self, cfg):
super(EmbNNRanker, self).__init__(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.normgrad = cfg.get('normgrad', False)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
# 'emb' = embeddings for both, 'emb_trees' = embeddings for tree only, 1-hot DA
# 'emb_tree', 'emb_prev' = tree-only embeddings
self.da_embs = cfg.get('nn', 'emb') == 'emb'
self.tree_embs = TreeEmbeddingExtract(cfg)
if self.da_embs:
self.da_embs = DAEmbeddingExtract(cfg)
else:
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.vectorizer = None
class RerankingClassifier(TFModel):
"""A classifier for trees that decides which DAIs are currently represented
(to be used in limiting candidate generator and/or re-scoring the trees)."""
def __init__(self, cfg):
super(RerankingClassifier, self).__init__(scope_name='rerank-' +
cfg.get('scope_suffix', ''))
self.cfg = cfg
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.tree_embs = cfg.get('nn', '').startswith('emb')
if self.tree_embs:
self.tree_embs = TreeEmbeddingClassifExtract(cfg)
self.emb_size = cfg.get('emb_size', 50)
self.mode = cfg.get('mode',
'tokens' if cfg.get('use_tokens') else 'trees')
self.nn_shape = cfg.get('nn_shape', 'ff')
self.num_hidden_units = cfg.get('num_hidden_units', 512)
self.passes = cfg.get('passes', 200)
self.min_passes = cfg.get('min_passes', 0)
self.alpha = cfg.get('alpha', 0.1)
self.randomize = cfg.get('randomize', True)
self.batch_size = cfg.get('batch_size', 1)
self.validation_freq = cfg.get('validation_freq', 10)
self.max_cores = cfg.get('max_cores')
self.cur_da = None
self.cur_da_bin = None
self.checkpoint_path = None
self.delex_slots = cfg.get('delex_slots', None)
if self.delex_slots:
self.delex_slots = set(self.delex_slots.split(','))
# Train Summaries
self.train_summary_dir = cfg.get('tb_summary_dir', None)
if self.train_summary_dir:
self.loss_summary_reranker = None
self.train_summary_op = None
self.train_summary_writer = None
def save_to_file(self, model_fname):
"""Save the classifier to a file (actually two files, one for configuration and one
for the TensorFlow graph, which must be stored separately).
@param model_fname: file name (for the configuration file); TF graph will be stored with a \
different extension
"""
model_fname = self.tf_check_filename(model_fname)
log_info("Saving classifier to %s..." % model_fname)
with file_stream(model_fname, 'wb', encoding=None) as fh:
pickle.dump(self.get_all_settings(),
fh,
protocol=pickle.HIGHEST_PROTOCOL)
tf_session_fname = re.sub(r'(.pickle)?(.gz)?$', '.tfsess', model_fname)
if hasattr(self, 'checkpoint_path') and self.checkpoint_path:
self.restore_checkpoint()
shutil.rmtree(os.path.dirname(self.checkpoint_path))
self.saver.save(self.session, tf_session_fname)
def get_all_settings(self):
"""Get all settings except the trained model parameters (to be stored in a pickle)."""
data = {
'cfg': self.cfg,
'da_feats': self.da_feats,
'da_vect': self.da_vect,
'tree_embs': self.tree_embs,
'input_shape': self.input_shape,
'num_outputs': self.num_outputs,
}
if self.tree_embs:
data['dict_size'] = self.dict_size
else:
data['tree_feats'] = self.tree_feats
data['tree_vect'] = self.tree_vect
return data
def _save_checkpoint(self):
"""Save a checkpoint to a temporary path; set `self.checkpoint_path` to the path
where it is saved; if called repeatedly, will always overwrite the last checkpoint."""
if not self.checkpoint_path:
path = tempfile.mkdtemp(suffix="", prefix="tftreecl-")
self.checkpoint_path = os.path.join(path, "ckpt")
log_info('Saving checkpoint to %s' % self.checkpoint_path)
self.saver.save(self.session, self.checkpoint_path)
def restore_checkpoint(self):
if not self.checkpoint_path:
return
self.saver.restore(self.session, self.checkpoint_path)
@staticmethod
def load_from_file(model_fname):
"""Load the reranker from a file (actually two files, one for configuration and one
for the TensorFlow graph, which must be stored separately).
@param model_fname: file name (for the configuration file); TF graph must be stored with a \
different extension
"""
log_info("Loading reranker from %s..." % model_fname)
with file_stream(model_fname, 'rb', encoding=None) as fh:
data = pickle.load(fh)
ret = RerankingClassifier(cfg=data['cfg'])
ret.load_all_settings(data)
# re-build TF graph and restore the TF session
tf_session_fname = os.path.abspath(
re.sub(r'(.pickle)?(.gz)?$', '.tfsess', model_fname))
ret._init_neural_network()
ret.saver.restore(ret.session, tf_session_fname)
return ret
def train(self,
das,
trees,
data_portion=1.0,
valid_das=None,
valid_trees=None):
"""Run training on the given training data.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (defaults to 1.0)
@param valid_das: validation data DAs
@param valid_trees: list of lists of corresponding paraphrases (same length as valid_das)
"""
log_info('Training reranking classifier...')
# initialize training
self._init_training(das, trees, data_portion)
if self.mode in ['tokens', 'tagged_lemmas'
] and valid_trees is not None:
valid_trees = [
self._tokens_to_flat_trees(
paraphrases, use_tags=self.mode == 'tagged_lemmas')
for paraphrases in valid_trees
]
# start training
top_comb_cost = float('nan')
for iter_no in xrange(1, self.passes + 1):
self.train_order = range(len(self.train_trees))
if self.randomize:
rnd.shuffle(self.train_order)
pass_cost, pass_diff = self._training_pass(iter_no)
if self.validation_freq and iter_no > self.min_passes and iter_no % self.validation_freq == 0:
valid_diff = 0
if valid_das:
valid_diff = np.sum([
np.sum(self.dist_to_da(d, t))
for d, t in zip(valid_das, valid_trees)
])
# cost combining validation and training data performance
# (+ "real" cost with negligible weight)
comb_cost = 1000 * valid_diff + 100 * pass_diff + pass_cost
log_info('Combined validation cost: %8.3f' % comb_cost)
# if we have the best model so far, save it as a checkpoint (overwrite previous)
if math.isnan(top_comb_cost) or comb_cost < top_comb_cost:
top_comb_cost = comb_cost
self._save_checkpoint()
# restore last checkpoint (best performance on devel data)
self.restore_checkpoint()
def classify(self, trees):
"""Classify the tree -- get DA slot-value pairs and DA type to which the tree
corresponds (as 1/0 array).
"""
if self.tree_embs:
inputs = np.array(
[self.tree_embs.get_embeddings(tree) for tree in trees])
else:
inputs = self.tree_vect.transform(
[self.tree_feats.get_features(tree, {}) for tree in trees])
fd = {}
self._add_inputs_to_feed_dict(inputs, fd)
results = self.session.run(self.outputs, feed_dict=fd)
# normalize & binarize the result
return np.array([[1. if r > 0 else 0. for r in result]
for result in results])
def _normalize_da(self, da):
if isinstance(da,
tuple): # if DA is actually context + DA, ignore context
da = da[1]
if self.delex_slots: # delexicalize the DA if needed
da = da.get_delexicalized(self.delex_slots)
return da
def init_run(self, da):
"""Remember the current DA for subsequent runs of `dist_to_cur_da`."""
self.cur_da = self._normalize_da(da)
da_bin = self.da_vect.transform(
[self.da_feats.get_features(None, {'da': self.cur_da})])[0]
self.cur_da_bin = da_bin != 0
def dist_to_da(self, da, trees):
"""Return Hamming distance of given trees to the given DA.
@param da: the DA as the base of the Hamming distance measure
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree
"""
da = self._normalize_da(da)
da_bin = self.da_vect.transform(
[self.da_feats.get_features(None, {'da': da})])[0]
da_bin = da_bin != 0
covered = self.classify(trees)
return [sum(abs(c - da_bin)) for c in covered]
def dist_to_cur_da(self, trees):
"""Return Hamming distance of given trees to the current DA (set in `init_run`).
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree
"""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [sum(abs(c - da_bin)) for c in covered]
def _init_training(self, das, trees, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (0.0-1.0)
"""
# read input from files or take it directly from parameters
if not isinstance(das, list):
log_info('Reading DAs from ' + das + '...')
das = read_das(das)
if not isinstance(trees, list):
log_info('Reading t-trees from ' + trees + '...')
ttree_doc = read_ttrees(trees)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language,
self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language,
self.selector)
trees = self._tokens_to_flat_trees(tls, use_tags=True)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
elif self.mode in ['tokens', 'tagged_lemmas']:
trees = self._tokens_to_flat_trees(
trees, use_tags=self.mode == 'tagged_lemmas')
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# ignore contexts, if they are contained in the DAs
if isinstance(self.train_das[0], tuple):
self.train_das = [da for (context, da) in self.train_das]
# delexicalize if DAs are lexicalized and we don't want that
if self.delex_slots:
self.train_das = [
da.get_delexicalized(self.delex_slots) for da in self.train_das
]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([
self.tree_embs.get_embeddings(tree)
for tree in self.train_trees
])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False,
binarize_numeric=True)
self.X = [
self.tree_feats.get_features(tree, {})
for tree in self.train_trees
]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [
self.da_feats.get_features(None, {'da': da})
for da in self.train_das
]
self.y = self.da_vect.fit_transform(self.y)
log_info('Number of binary classes: %d.' %
len(self.da_vect.get_feature_names()))
# initialize I/O shapes
if not self.tree_embs:
self.input_shape = list(self.X[0].shape)
else:
self.input_shape = self.tree_embs.get_embeddings_shape()
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
# initialize the NN variables
self.session.run(tf.global_variables_initializer())
def _tokens_to_flat_trees(self, sents, use_tags=False):
"""Use sentences (pairs token-tag) read from Treex files and convert them into flat
trees (each token has a node right under the root, lemma is the token, formeme is 'x').
Uses TokenEmbeddingSeq2SeqExtract conversion there and back.
@param sents: sentences to be converted
@param use_tags: use tags in the embeddings? (only for lemma-tag pairs in training, \
not testing)
@return: a list of flat trees
"""
tree_embs = (TokenEmbeddingSeq2SeqExtract(cfg=self.cfg) if not use_tags
else TaggedLemmasEmbeddingSeq2SeqExtract(cfg=self.cfg))
tree_embs.init_dict(sents)
# no postprocessing, i.e. keep lowercasing/plural splitting if set in the configuration
return [
tree_embs.ids_to_tree(tree_embs.get_embeddings(sent),
postprocess=False) for sent in sents
]
def _init_neural_network(self):
"""Create the neural network for classification, according to the self.nn_shape
parameter (as set in configuration)."""
# set TensorFlow random seed
tf.set_random_seed(rnd.randint(-sys.maxint, sys.maxint))
self.targets = tf.placeholder(tf.float32, [None, self.num_outputs],
name='targets')
with tf.variable_scope(self.scope_name):
# feedforward networks
if self.nn_shape.startswith('ff'):
self.inputs = tf.placeholder(tf.float32,
[None] + self.input_shape,
name='inputs')
num_ff_layers = 2
if self.nn_shape[-1] in ['0', '1', '3', '4']:
num_ff_layers = int(self.nn_shape[-1])
self.outputs = self._ff_layers('ff', num_ff_layers,
self.inputs)
# RNNs
elif self.nn_shape.startswith('rnn'):
self.initial_state = tf.placeholder(tf.float32,
[None, self.emb_size])
self.inputs = [
tf.placeholder(tf.int32, [None], name=('enc_inp-%d' % i))
for i in xrange(self.input_shape[0])
]
self.cell = tf.contrib.rnn.BasicLSTMCell(self.emb_size)
self.outputs = self._rnn('rnn', self.inputs)
# the cost as computed by TF actually adds a "fake" sigmoid layer on top
# (or is computed as if there were a sigmoid layer on top)
self.cost = tf.reduce_mean(
tf.reduce_sum(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.outputs,
labels=self.targets,
name='CE'), 1))
# NB: this would have been the "true" cost function, if there were a "real" sigmoid layer on top.
# However, it is not numerically stable in practice, so we have to use the TF function.
# self.cost = tf.reduce_mean(tf.reduce_sum(self.targets * -tf.log(self.outputs)
# + (1 - self.targets) * -tf.log(1 - self.outputs), 1))
self.optimizer = tf.train.AdamOptimizer(self.alpha)
self.train_func = self.optimizer.minimize(self.cost)
# Tensorboard summaries
if self.train_summary_dir:
self.loss_summary_reranker = tf.summary.scalar(
"loss_reranker", self.cost)
self.train_summary_op = tf.summary.merge(
[self.loss_summary_reranker])
# initialize session
session_config = None
if self.max_cores:
session_config = tf.ConfigProto(
inter_op_parallelism_threads=self.max_cores,
intra_op_parallelism_threads=self.max_cores)
self.session = tf.Session(config=session_config)
# this helps us load/save the model
self.saver = tf.train.Saver(tf.global_variables())
if self.train_summary_dir: # Tensorboard summary writer
self.train_summary_writer = tf.summary.FileWriter(
os.path.join(self.train_summary_dir, "reranker"),
self.session.graph)
def _ff_layers(self, name, num_layers, X):
width = [np.prod(self.input_shape)] + (
num_layers * [self.num_hidden_units]) + [self.num_outputs]
# the last layer should be a sigmoid, but TF simulates it for us in cost computation
# so the output is "unnormalized sigmoids"
activ = (num_layers * [tf.tanh]) + [tf.identity]
Y = X
for i in xrange(num_layers + 1):
w = tf.get_variable(
name + ('-w%d' % i), (width[i], width[i + 1]),
initializer=tf.random_normal_initializer(stddev=0.1))
b = tf.get_variable(name + ('-b%d' % i), (width[i + 1], ),
initializer=tf.constant_initializer())
Y = activ[i](tf.matmul(Y, w) + b)
return Y
def _rnn(self, name, enc_inputs):
encoder_cell = tf.contrib.rnn.EmbeddingWrapper(self.cell,
self.dict_size,
self.emb_size)
encoder_outputs, encoder_state = tf.contrib.rnn.static_rnn(
encoder_cell, enc_inputs, dtype=tf.float32)
# TODO for historical reasons, the last layer uses both output and state.
# try this just with outputs (might work exactly the same)
if isinstance(self.cell.state_size, tf.contrib.rnn.LSTMStateTuple):
state_size = self.cell.state_size.c + self.cell.state_size.h
final_input = tf.concat(axis=1,
values=encoder_state) # concat c + h
else:
state_size = self.cell.state_size
final_input = encoder_state
w = tf.get_variable(
name + '-w', (state_size, self.num_outputs),
initializer=tf.random_normal_initializer(stddev=0.1))
b = tf.get_variable(name + 'b', (self.num_outputs, ),
initializer=tf.constant_initializer())
return tf.matmul(final_input, w) + b
def _batches(self):
"""Create batches from the input; use as iterator."""
for i in xrange(0, len(self.train_order), self.batch_size):
yield self.train_order[i:i + self.batch_size]
def _add_inputs_to_feed_dict(self, inputs, fd):
if self.nn_shape.startswith('rnn'):
fd[self.initial_state] = np.zeros([inputs.shape[0], self.emb_size])
sliced_inputs = np.squeeze(np.array(
np.split(np.array([ex for ex in inputs if ex is not None]),
len(inputs[0]),
axis=1)),
axis=2)
for input_, slice_ in zip(self.inputs, sliced_inputs):
fd[input_] = slice_
else:
fd[self.inputs] = inputs
def _training_pass(self, pass_no):
"""Perform one training pass through the whole training data, print statistics."""
pass_start_time = time.time()
log_debug('\n***\nTR %05d:' % pass_no)
log_debug("Train order: " + str(self.train_order))
pass_cost = 0
pass_diff = 0
for tree_nos in self._batches():
log_debug('TREE-NOS: ' + str(tree_nos))
log_debug("\n".join(
unicode(self.train_trees[i]) + "\n" +
unicode(self.train_das[i]) for i in tree_nos))
log_debug('Y: ' + str(self.y[tree_nos]))
fd = {self.targets: self.y[tree_nos]}
self._add_inputs_to_feed_dict(self.X[tree_nos], fd)
if self.train_summary_dir: # also compute Tensorboard summaries
results, cost, _, train_summary_op = self.session.run(
[
self.outputs, self.cost, self.train_func,
self.train_summary_op
],
feed_dict=fd)
else:
results, cost, _ = self.session.run(
[self.outputs, self.cost, self.train_func], feed_dict=fd)
bin_result = np.array([[1. if r > 0 else 0. for r in result]
for result in results])
log_debug('R: ' + str(bin_result))
log_debug('COST: %f' % cost)
log_debug('DIFF: %d' %
np.sum(np.abs(self.y[tree_nos] - bin_result)))
pass_cost += cost
pass_diff += np.sum(np.abs(self.y[tree_nos] - bin_result))
# print and return statistics
self._print_pass_stats(
pass_no,
datetime.timedelta(seconds=(time.time() - pass_start_time)),
pass_cost, pass_diff)
if self.train_summary_dir: # Tensorboard: iteration summary
self.train_summary_writer.add_summary(train_summary_op, pass_no)
return pass_cost, pass_diff
def _print_pass_stats(self, pass_no, time, cost, diff):
log_info('PASS %03d: duration %s, cost %f, diff %d' %
(pass_no, str(time), cost, diff))
def evaluate_file(self, das_file, ttree_file):
"""Evaluate the reranking classifier on a given pair of DA/tree files (show the
total Hamming distance and total number of DAIs)
@param das_file: DA file path
@param ttree_file: trees/sentences file path
@return: a tuple (total DAIs, distance)
"""
das = read_das(das_file)
ttree_doc = read_ttrees(ttree_file)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language,
self.selector)
trees = self._tokens_to_flat_trees(tls)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
da_len = 0
dist = 0
for da, tree in zip(das, trees):
da_len += len(da)
dist += self.dist_to_da(da, [tree])[0]
return da_len, dist
def _init_training(self, das, trees, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (0.0-1.0)
"""
# read input from files or take it directly from parameters
if not isinstance(das, list):
log_info('Reading DAs from ' + das + '...')
das = read_das(das)
if not isinstance(trees, list):
log_info('Reading t-trees from ' + trees + '...')
ttree_doc = read_ttrees(trees)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language,
self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language,
self.selector)
trees = self._tokens_to_flat_trees(tls, use_tags=True)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
elif self.mode in ['tokens', 'tagged_lemmas']:
trees = self._tokens_to_flat_trees(
trees, use_tags=self.mode == 'tagged_lemmas')
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# ignore contexts, if they are contained in the DAs
if isinstance(self.train_das[0], tuple):
self.train_das = [da for (context, da) in self.train_das]
# delexicalize if DAs are lexicalized and we don't want that
if self.delex_slots:
self.train_das = [
da.get_delexicalized(self.delex_slots) for da in self.train_das
]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([
self.tree_embs.get_embeddings(tree)
for tree in self.train_trees
])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False,
binarize_numeric=True)
self.X = [
self.tree_feats.get_features(tree, {})
for tree in self.train_trees
]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [
self.da_feats.get_features(None, {'da': da})
for da in self.train_das
]
self.y = self.da_vect.fit_transform(self.y)
log_info('Number of binary classes: %d.' %
len(self.da_vect.get_feature_names()))
# initialize I/O shapes
if not self.tree_embs:
self.input_shape = list(self.X[0].shape)
else:
self.input_shape = self.tree_embs.get_embeddings_shape()
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
# initialize the NN variables
self.session.run(tf.global_variables_initializer())
class TreeClassifier(object):
"""A classifier for trees that decides which DAIs are currently represented
(to be used in limiting candidate generator and/or re-scoring the trees)."""
def __init__(self, cfg):
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.tree_embs = cfg.get('nn', '').startswith('emb')
if self.tree_embs:
self.tree_embs = TreeEmbeddingExtract(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.num_hidden_units = cfg.get('num_hidden_units', 512)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
self.init = cfg.get('initialization', 'uniform_glorot10')
self.passes = cfg.get('passes', 200)
self.alpha = cfg.get('alpha', 0.1)
self.randomize = cfg.get('randomize', True)
self.batch_size = cfg.get('batch_size', 1)
self.cur_da = None
self.cur_da_bin = None
@staticmethod
def load_from_file(fname):
log_info('Loading model from ' + fname)
with file_stream(fname, mode='rb', encoding=None) as fh:
classif = pickle.load(fh)
return classif
def save_to_file(self, fname):
log_info('Saving model to ' + fname)
with file_stream(fname, mode='wb', encoding=None) as fh:
pickle.dump(self, fh, pickle.HIGHEST_PROTOCOL)
def train(self, das_file, ttree_file, data_portion=1.0):
"""Run training on the given training data."""
self._init_training(das_file, ttree_file, data_portion)
for iter_no in xrange(1, self.passes + 1):
self.train_order = range(len(self.train_trees))
if self.randomize:
rnd.shuffle(self.train_order)
self._training_pass(iter_no)
def classify(self, trees):
"""Classify the tree -- get DA slot-value pairs and DA type to which the tree
corresponds (as 1/0 array).
This does not have a lot of practical use here, see is_subset_of_da.
"""
if self.tree_embs:
X = np.array(
[self.tree_embs.get_embeddings(tree) for tree in trees])
else:
X = self.tree_vect.transform(
[self.tree_feats.get_features(tree, {}) for tree in trees])
# binarize the result
return np.array([[1. if r > 0.5 else 0. for r in result]
for result in self.classif.classif(X)])
def is_subset_of_da(self, da, trees):
"""Given a DA and an array of trees, this gives a boolean array indicating which
trees currently cover/describe a subset of the DA.
@param da: the input DA against which the trees should be tested
@param trees: the trees to test against the DA
@return: boolean array, with True where the tree covers/describes a subset of the DA
"""
# get 1-hot representation of the DA
da_bin = self.da_vect.transform(
[self.da_feats.get_features(None, {'da': da})])[0]
# convert it to array of booleans
da_bin = da_bin != 0
# classify the trees
covered = self.classify(trees)
# decide whether 1's in their 1-hot vectors are subsets of True's in da_bin
return [((c != 0) | da_bin == da_bin).all() for c in covered]
def init_run(self, da):
"""Remember the current DA for subsequent runs of `is_subset_of_cur_da`."""
self.cur_da = da
da_bin = self.da_vect.transform(
[self.da_feats.get_features(None, {'da': da})])[0]
self.cur_da_bin = da_bin != 0
def is_subset_of_cur_da(self, trees):
"""Same as `is_subset_of_da`, but using `self.cur_da` set via `init_run`."""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [((c != 0) | da_bin == da_bin).all() for c in covered]
def corresponds_to_cur_da(self, trees):
"""Given an array of trees, this gives a boolean array indicating which
trees currently cover exactly the current DA (set via `init_run`).
@param trees: the trees to test against the current DA
@return: boolean array, with True where the tree covers/describes a subset of the current DA
"""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [((c != 0) == da_bin).all() for c in covered]
def _init_training(self, das_file, ttree_file, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
"""
# read input
log_info('Reading DAs from ' + das_file + '...')
das = read_das(das_file)
log_info('Reading t-trees from ' + ttree_file + '...')
ttree_doc = read_ttrees(ttree_file)
trees = trees_from_doc(ttree_doc, self.language, self.selector)
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([
self.tree_embs.get_embeddings(tree)
for tree in self.train_trees
])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False,
binarize_numeric=True)
self.X = [
self.tree_feats.get_features(tree, {})
for tree in self.train_trees
]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [
self.da_feats.get_features(None, {'da': da})
for da in self.train_das
]
self.y = self.da_vect.fit_transform(self.y)
# initialize I/O shapes
self.input_shape = [list(self.X[0].shape)]
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
def _init_neural_network(self):
"""Create the neural network for classification, according to the self.nn_shape
parameter (as set in configuration)."""
layers = []
if self.tree_embs:
layers.append([
Embedding('emb', self.dict_size, self.emb_size, 'uniform_005')
])
# feedforward networks
if self.nn_shape.startswith('ff'):
if self.tree_embs:
layers.append([Flatten('flat')])
num_ff_layers = 2
if self.nn_shape[-1] in ['0', '1', '3', '4']:
num_ff_layers = int(self.nn_shape[-1])
layers += self._ff_layers('ff', num_ff_layers)
# convolutional networks
elif 'conv' in self.nn_shape or 'pool' in self.nn_shape:
assert self.tree_embs # convolution makes no sense without embeddings
num_conv = 0
if 'conv' in self.nn_shape:
num_conv = 1
if 'conv2' in self.nn_shape:
num_conv = 2
pooling = None
if 'maxpool' in self.nn_shape:
pooling = T.max
elif 'avgpool' in self.nn_shape:
pooling = T.mean
layers += self._conv_layers('conv', num_conv, pooling)
layers.append([Flatten('flat')])
layers += self._ff_layers('ff', 1)
# input types: integer 3D for tree embeddings (batch + 2D embeddings),
# float 2D (matrix) for binary input (batch + features)
input_types = (T.itensor3, ) if self.tree_embs else (T.fmatrix, )
# create the network, connect layers
self.classif = ClassifNN(layers,
self.input_shape,
input_types,
normgrad=False)
log_info("Network shape:\n\n" + str(self.classif))
def _ff_layers(self, name, num_layers):
ret = []
for i in xrange(num_layers):
ret.append([
FeedForward(name + str(i + 1), self.num_hidden_units, T.tanh,
self.init)
])
ret.append([
FeedForward('output', self.num_outputs, T.nnet.sigmoid, self.init)
])
return ret
def _conv_layers(self, name, num_layers=1, pooling=None):
ret = []
for i in xrange(num_layers):
ret.append([
Conv1D(name + str(i + 1),
filter_length=self.cnn_filter_length,
num_filters=self.cnn_num_filters,
init=self.init,
activation=T.tanh)
])
if pooling is not None:
ret.append(
[Pool1D(name + str(i + 1) + 'pool', pooling_func=pooling)])
return ret
def batches(self):
for i in xrange(0, len(self.train_order), self.batch_size):
yield self.train_order[i:i + self.batch_size]
def _training_pass(self, pass_no):
"""Perform one training pass through the whole training data, print statistics."""
pass_start_time = time.time()
log_debug('\n***\nTR %05d:' % pass_no)
log_debug("Train order: " + str(self.train_order))
pass_cost = 0
pass_diff = 0
for tree_nos in self.batches():
log_debug('TREE-NOS: ' + str(tree_nos))
log_debug("\n".join(
unicode(self.train_trees[i]) + "\n" +
unicode(self.train_das[i]) for i in tree_nos))
log_debug('Y: ' + str(self.y[tree_nos]))
results = self.classif.classif(self.X[tree_nos])
cost_gcost = self.classif.update(self.X[tree_nos],
self.y[tree_nos], self.alpha)
bin_result = np.array([[1. if r > 0.5 else 0. for r in result]
for result in results])
log_debug('R: ' + str(bin_result))
log_debug('COST: %f' % cost_gcost[0])
log_debug('DIFF: %d' %
np.sum(np.abs(self.y[tree_nos] - bin_result)))
pass_cost += cost_gcost[0]
pass_diff += np.sum(np.abs(self.y[tree_nos] - bin_result))
# print and return statistics
self._print_pass_stats(
pass_no,
datetime.timedelta(seconds=(time.time() - pass_start_time)),
pass_cost, pass_diff)
def _print_pass_stats(self, pass_no, time, cost, diff):
log_info('PASS %03d: duration %s, cost %f, diff %d' %
(pass_no, str(time), cost, diff))
class EmbNNRanker(NNRanker):
"""A ranker using MR and tree embeddings in a NN."""
def __init__(self, cfg):
super(EmbNNRanker, self).__init__(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.normgrad = cfg.get('normgrad', False)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
# 'emb' = embeddings for both, 'emb_trees' = embeddings for tree only, 1-hot DA
# 'emb_tree', 'emb_prev' = tree-only embeddings
self.da_embs = cfg.get('nn', 'emb') == 'emb'
self.tree_embs = TreeEmbeddingExtract(cfg)
if self.da_embs:
self.da_embs = DAEmbeddingExtract(cfg)
else:
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.vectorizer = None
def _init_training(self, das_file, ttree_file, data_portion):
super(EmbNNRanker, self)._init_training(das_file, ttree_file, data_portion)
self._init_dict()
self._init_neural_network()
self.train_feats = [self._extract_feats(tree, da)
for tree, da in zip(self.train_trees, self.train_das)]
self.w_after_iter = []
self.update_weights_sum()
def _init_dict(self):
"""Initialize word -> integer dictionaries, starting from a minimum
valid value, always adding a new integer to unknown values to prevent
clashes among different types of inputs."""
# avoid dictionary clashes between DAs and tree embeddings
# – remember current highest index number
dict_ord = None
# DA embeddings
if self.da_embs:
dict_ord = self.da_embs.init_dict(self.train_das)
# DA one-hot representation
else:
X = []
for da, tree in zip(self.train_das, self.train_trees):
X.append(self.da_feats.get_features(tree, {'da': da}))
self.vectorizer = DictVectorizer(sparse=False, binarize_numeric=True)
self.vectorizer.fit(X)
# tree embeddings
# remember last dictionary key to initialize embeddings with enough rows
self.dict_size = self.tree_embs.init_dict(self.train_trees, dict_ord)
def _score(self, cand_embs):
return self.nn.score([cand_embs[0]], [cand_embs[1]])[0]
def _extract_feats(self, tree, da):
"""Extract DA and tree embeddings (return as a pair)."""
if self.da_embs:
# DA embeddings
da_repr = self.da_embs.get_embeddings(da)
else:
# DA one-hot representation
da_repr = self.vectorizer.transform([self.da_feats.get_features(tree, {'da': da})])[0]
# tree embeddings
tree_emb_idxs = self.tree_embs.get_embeddings(tree)
return (da_repr, tree_emb_idxs)
def _init_neural_network(self):
# initial layer – tree embeddings & DA 1-hot or embeddings
# input shapes don't contain the batch dimension, but the input Theano types do!
if self.da_embs:
input_shapes = (self.da_embs.get_embeddings_shape(),
self.tree_embs.get_embeddings_shape())
input_types = (T.itensor3, T.itensor3)
layers = [[Embedding('emb_da', self.dict_size, self.emb_size, 'uniform_005'),
Embedding('emb_tree', self.dict_size, self.emb_size, 'uniform_005')]]
else:
input_shapes = ([len(self.vectorizer.get_feature_names())],
self.tree_embs.get_embeddings_shape())
input_types = (T.fmatrix, T.itensor3)
layers = [[Identity('id_da'),
Embedding('emb_tree', self.dict_size, self.emb_size, 'uniform_005')]]
# plain feed-forward networks
if self.nn_shape.startswith('ff'):
layers += [[Flatten('flat_da'), Flatten('flat_tree')], [Concat('concat')]]
num_ff_layers = 2
if self.nn_shape[-1] in ['3', '4']:
num_ff_layers = int(self.nn_shape[-1])
layers += self._ff_layers('ff', num_ff_layers, perc_layer=True)
# convolution with or without max/avg-pooling
elif self.nn_shape.startswith('conv'):
num_conv_layers = 2 if self.nn_shape.startswith('conv2') else 1
pooling = None
if 'maxpool' in self.nn_shape:
pooling = T.max
elif 'avgpool' in self.nn_shape:
pooling = T.mean
if self.da_embs:
da_layers = self._conv_layers('conv_da', num_conv_layers, pooling=pooling)
else:
da_layers = self._id_layers('id_da',
num_conv_layers + (1 if pooling is not None else 0))
tree_layers = self._conv_layers('conv_tree', num_conv_layers, pooling=pooling)
for da_layer, tree_layer in zip(da_layers, tree_layers):
layers.append([da_layer[0], tree_layer[0]])
layers += [[Flatten('flat_da'), Flatten('flat_tree')], [Concat('concat')]]
layers += self._ff_layers('ff', 2, perc_layer=True)
# max-pooling without convolution
elif 'maxpool-ff' in self.nn_shape:
layers += [[Pool1D('mp_da') if self.da_embs else Identity('id_da'),
Pool1D('mp_trees')]
[Concat('concat')], [Flatten('flat')]]
layers += self._ff_layers('ff', 2, perc_layer=True),
# dot-product FF network
elif 'dot' in self.nn_shape:
# with max or average pooling
if 'maxpool' in self.nn_shape or 'avgpool' in self.nn_shape:
pooling = T.mean if 'avgpool' in self.nn_shape else T.max
layers += [[Pool1D('mp_da', pooling_func=pooling)
if self.da_embs else Identity('id_da'),
Pool1D('mp_tree', pooling_func=pooling)]]
layers += [[Flatten('flat_da') if self.da_embs else Identity('id_da'),
Flatten('flat_tree')]]
num_ff_layers = int(self.nn_shape[-1]) if self.nn_shape[-1] in ['2', '3', '4'] else 1
for da_layer, tree_layer in zip(self._ff_layers('ff_da', num_ff_layers),
self._ff_layers('ff_tree', num_ff_layers)):
layers.append([da_layer[0], tree_layer[0]])
layers.append([DotProduct('dot')])
# input: batch * word * sub-embeddings
self.nn = RankNN(layers, input_shapes, input_types, self.normgrad)
log_info("Network shape:\n\n" + str(self.nn))
def _conv_layers(self, name, num_layers=1, pooling=None):
ret = []
for i in range(num_layers):
ret.append([Conv1D(name + str(i + 1),
filter_length=self.cnn_filter_length,
num_filters=self.cnn_num_filters,
init=self.init, activation=T.tanh)])
if pooling is not None:
ret.append([Pool1D(name + str(i + 1) + 'pool', pooling_func=pooling)])
return ret
def _id_layers(self, name, num_layers):
ret = []
for i in range(num_layers):
ret.append([Identity(name + str(i + 1))])
return ret
def _update_nn(self, bad_feats, good_feats, rate):
"""Changing the NN update call to support arrays of parameters."""
# TODO: this is just adding another dimension to fit the parallelized scoring
# (even if updates are not parallelized). Make it nicer.
bad_feats = ([bad_feats[0]], [bad_feats[1]])
good_feats = ([good_feats[0]], [good_feats[1]])
cost_gcost = self.nn.update(*(bad_feats + good_feats + (rate,)))
log_debug('Cost:' + str(cost_gcost[0]))
param_vals = [param.get_value() for param in self.nn.params]
log_debug('Param norms : ' + str(self._l2s(param_vals)))
log_debug('Gparam norms: ' + str(self._l2s(cost_gcost[1:])))
def _embs_to_str(self):
out = ""
da_emb = self.nn.layers[0][0].e.get_value()
tree_emb = self.nn.layers[0][1].e.get_value()
for idx, emb in enumerate(da_emb):
for key, val in list(self.dict_slot.items()):
if val == idx:
out += key + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for key, val in list(self.dict_value.items()):
if val == idx:
out += key + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for idx, emb in enumerate(tree_emb):
for key, val in list(self.dict_t_lemma.items()):
if val == idx:
out += str(key) + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for key, val in list(self.dict_formeme.items()):
if val == idx:
out += str(key) + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
return out
def _l2s(self, params):
"""Compute L2-norm of all members of the given list."""
return [np.linalg.norm(param) for param in params]
def store_iter_weights(self):
"""Remember the current weights to be used for averaged perceptron."""
# fh = open('embs.txt', 'a')
# print >> fh, '---', self._embs_to_str()
# fh.close()
self.w_after_iter.append(self.nn.get_param_values())
def score_all(self, trees, da):
cand_embs = [self._extract_feats(tree, da) for tree in trees]
score = self.nn.score([emb[0] for emb in cand_embs], [emb[1] for emb in cand_embs])
return np.atleast_1d(score[0])
from flect.config import Config
from tgen.features import Features
from tgen.futil import trees_from_doc, read_ttrees, read_das
import sys
import timeit
import datetime
if len(sys.argv[1:]) != 3:
sys.exit('Usage: ./bench_feats.py features_cfg.py trees.yaml.gz das.txt')
print >> sys.stderr, 'Loading...'
cfg = Config(sys.argv[1])
trees = trees_from_doc(read_ttrees(sys.argv[2]), 'en', '')
das = read_das(sys.argv[3])
feats = Features(cfg['features'])
def test_func():
for tree, da in zip(trees, das):
feats.get_features(tree, {'da': da})
print >> sys.stderr, 'Running test...'
secs = timeit.timeit('test_func()',
setup='from __main__ import test_func',
number=10)
td = datetime.timedelta(seconds=secs)
print >> sys.stderr, 'Time taken: %s' % str(td)
class TreeClassifier(object):
"""A classifier for trees that decides which DAIs are currently represented
(to be used in limiting candidate generator and/or re-scoring the trees)."""
def __init__(self, cfg):
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.tree_embs = cfg.get('nn', '').startswith('emb')
if self.tree_embs:
self.tree_embs = TreeEmbeddingExtract(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.num_hidden_units = cfg.get('num_hidden_units', 512)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
self.init = cfg.get('initialization', 'uniform_glorot10')
self.passes = cfg.get('passes', 200)
self.alpha = cfg.get('alpha', 0.1)
self.randomize = cfg.get('randomize', True)
self.batch_size = cfg.get('batch_size', 1)
self.cur_da = None
self.cur_da_bin = None
@staticmethod
def load_from_file(fname):
log_info('Loading model from ' + fname)
with file_stream(fname, mode='rb', encoding=None) as fh:
classif = pickle.load(fh)
return classif
def save_to_file(self, fname):
log_info('Saving model to ' + fname)
with file_stream(fname, mode='wb', encoding=None) as fh:
pickle.dump(self, fh, pickle.HIGHEST_PROTOCOL)
def train(self, das_file, ttree_file, data_portion=1.0):
"""Run training on the given training data."""
self._init_training(das_file, ttree_file, data_portion)
for iter_no in xrange(1, self.passes + 1):
self.train_order = range(len(self.train_trees))
if self.randomize:
rnd.shuffle(self.train_order)
self._training_pass(iter_no)
def classify(self, trees):
"""Classify the tree -- get DA slot-value pairs and DA type to which the tree
corresponds (as 1/0 array).
This does not have a lot of practical use here, see is_subset_of_da.
"""
if self.tree_embs:
X = np.array([self.tree_embs.get_embeddings(tree) for tree in trees])
else:
X = self.tree_vect.transform([self.tree_feats.get_features(tree, {}) for tree in trees])
# binarize the result
return np.array([[1. if r > 0.5 else 0. for r in result]
for result in self.classif.classif(X)])
def is_subset_of_da(self, da, trees):
"""Given a DA and an array of trees, this gives a boolean array indicating which
trees currently cover/describe a subset of the DA.
@param da: the input DA against which the trees should be tested
@param trees: the trees to test against the DA
@return: boolean array, with True where the tree covers/describes a subset of the DA
"""
# get 1-hot representation of the DA
da_bin = self.da_vect.transform([self.da_feats.get_features(None, {'da': da})])[0]
# convert it to array of booleans
da_bin = da_bin != 0
# classify the trees
covered = self.classify(trees)
# decide whether 1's in their 1-hot vectors are subsets of True's in da_bin
return [((c != 0) | da_bin == da_bin).all() for c in covered]
def init_run(self, da):
"""Remember the current DA for subsequent runs of `is_subset_of_cur_da`."""
self.cur_da = da
da_bin = self.da_vect.transform([self.da_feats.get_features(None, {'da': da})])[0]
self.cur_da_bin = da_bin != 0
def is_subset_of_cur_da(self, trees):
"""Same as `is_subset_of_da`, but using `self.cur_da` set via `init_run`."""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [((c != 0) | da_bin == da_bin).all() for c in covered]
def corresponds_to_cur_da(self, trees):
"""Given an array of trees, this gives a boolean array indicating which
trees currently cover exactly the current DA (set via `init_run`).
@param trees: the trees to test against the current DA
@return: boolean array, with True where the tree covers/describes a subset of the current DA
"""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [((c != 0) == da_bin).all() for c in covered]
def _init_training(self, das_file, ttree_file, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
"""
# read input
log_info('Reading DAs from ' + das_file + '...')
das = read_das(das_file)
log_info('Reading t-trees from ' + ttree_file + '...')
ttree_doc = read_ttrees(ttree_file)
trees = trees_from_doc(ttree_doc, self.language, self.selector)
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DialogueAct()
empty_da.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([self.tree_embs.get_embeddings(tree) for tree in self.train_trees])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.X = [self.tree_feats.get_features(tree, {}) for tree in self.train_trees]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [self.da_feats.get_features(None, {'da': da}) for da in self.train_das]
self.y = self.da_vect.fit_transform(self.y)
# initialize I/O shapes
self.input_shape = [list(self.X[0].shape)]
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
def _init_neural_network(self):
"""Create the neural network for classification, according to the self.nn_shape
parameter (as set in configuration)."""
layers = []
if self.tree_embs:
layers.append([Embedding('emb', self.dict_size, self.emb_size, 'uniform_005')])
# feedforward networks
if self.nn_shape.startswith('ff'):
if self.tree_embs:
layers.append([Flatten('flat')])
num_ff_layers = 2
if self.nn_shape[-1] in ['0', '1', '3', '4']:
num_ff_layers = int(self.nn_shape[-1])
layers += self._ff_layers('ff', num_ff_layers)
# convolutional networks
elif 'conv' in self.nn_shape or 'pool' in self.nn_shape:
assert self.tree_embs # convolution makes no sense without embeddings
num_conv = 0
if 'conv' in self.nn_shape:
num_conv = 1
if 'conv2' in self.nn_shape:
num_conv = 2
pooling = None
if 'maxpool' in self.nn_shape:
pooling = T.max
elif 'avgpool' in self.nn_shape:
pooling = T.mean
layers += self._conv_layers('conv', num_conv, pooling)
layers.append([Flatten('flat')])
layers += self._ff_layers('ff', 1)
# input types: integer 3D for tree embeddings (batch + 2D embeddings),
# float 2D (matrix) for binary input (batch + features)
input_types = (T.itensor3,) if self.tree_embs else (T.fmatrix,)
# create the network, connect layers
self.classif = ClassifNN(layers, self.input_shape, input_types, normgrad=False)
log_info("Network shape:\n\n" + str(self.classif))
def _ff_layers(self, name, num_layers):
ret = []
for i in xrange(num_layers):
ret.append([FeedForward(name + str(i + 1), self.num_hidden_units, T.tanh, self.init)])
ret.append([FeedForward('output', self.num_outputs, T.nnet.sigmoid, self.init)])
return ret
def _conv_layers(self, name, num_layers=1, pooling=None):
ret = []
for i in xrange(num_layers):
ret.append([Conv1D(name + str(i + 1),
filter_length=self.cnn_filter_length,
num_filters=self.cnn_num_filters,
init=self.init, activation=T.tanh)])
if pooling is not None:
ret.append([Pool1D(name + str(i + 1) + 'pool', pooling_func=pooling)])
return ret
def batches(self):
for i in xrange(0, len(self.train_order), self.batch_size):
yield self.train_order[i: i + self.batch_size]
def _training_pass(self, pass_no):
"""Perform one training pass through the whole training data, print statistics."""
pass_start_time = time.time()
log_debug('\n***\nTR %05d:' % pass_no)
log_debug("Train order: " + str(self.train_order))
pass_cost = 0
pass_diff = 0
for tree_nos in self.batches():
log_debug('TREE-NOS: ' + str(tree_nos))
log_debug("\n".join(unicode(self.train_trees[i]) + "\n" + unicode(self.train_das[i])
for i in tree_nos))
log_debug('Y: ' + str(self.y[tree_nos]))
results = self.classif.classif(self.X[tree_nos])
cost_gcost = self.classif.update(self.X[tree_nos], self.y[tree_nos], self.alpha)
bin_result = np.array([[1. if r > 0.5 else 0. for r in result] for result in results])
log_debug('R: ' + str(bin_result))
log_debug('COST: %f' % cost_gcost[0])
log_debug('DIFF: %d' % np.sum(np.abs(self.y[tree_nos] - bin_result)))
pass_cost += cost_gcost[0]
pass_diff += np.sum(np.abs(self.y[tree_nos] - bin_result))
# print and return statistics
self._print_pass_stats(pass_no, datetime.timedelta(seconds=(time.time() - pass_start_time)),
pass_cost, pass_diff)
def _print_pass_stats(self, pass_no, time, cost, diff):
log_info('PASS %03d: duration %s, cost %f, diff %d' % (pass_no, str(time), cost, diff))
class EmbNNRanker(NNRanker):
"""A ranker using MR and tree embeddings in a NN."""
def __init__(self, cfg):
super(EmbNNRanker, self).__init__(cfg)
self.emb_size = cfg.get('emb_size', 20)
self.nn_shape = cfg.get('nn_shape', 'ff')
self.normgrad = cfg.get('normgrad', False)
self.cnn_num_filters = cfg.get('cnn_num_filters', 3)
self.cnn_filter_length = cfg.get('cnn_filter_length', 3)
# 'emb' = embeddings for both, 'emb_trees' = embeddings for tree only, 1-hot DA
# 'emb_tree', 'emb_prev' = tree-only embeddings
self.da_embs = cfg.get('nn', 'emb') == 'emb'
self.tree_embs = TreeEmbeddingExtract(cfg)
if self.da_embs:
self.da_embs = DAEmbeddingExtract(cfg)
else:
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.vectorizer = None
def _init_training(self, das_file, ttree_file, data_portion):
super(EmbNNRanker, self)._init_training(das_file, ttree_file, data_portion)
self._init_dict()
self._init_neural_network()
self.train_feats = [self._extract_feats(tree, da)
for tree, da in zip(self.train_trees, self.train_das)]
self.w_after_iter = []
self.update_weights_sum()
def _init_dict(self):
"""Initialize word -> integer dictionaries, starting from a minimum
valid value, always adding a new integer to unknown values to prevent
clashes among different types of inputs."""
# avoid dictionary clashes between DAs and tree embeddings
# – remember current highest index number
dict_ord = None
# DA embeddings
if self.da_embs:
dict_ord = self.da_embs.init_dict(self.train_das)
# DA one-hot representation
else:
X = []
for da, tree in zip(self.train_das, self.train_trees):
X.append(self.da_feats.get_features(tree, {'da': da}))
self.vectorizer = DictVectorizer(sparse=False, binarize_numeric=True)
self.vectorizer.fit(X)
# tree embeddings
# remember last dictionary key to initialize embeddings with enough rows
self.dict_size = self.tree_embs.init_dict(self.train_trees, dict_ord)
def _score(self, cand_embs):
return self.nn.score([cand_embs[0]], [cand_embs[1]])[0]
def _extract_feats(self, tree, da):
"""Extract DA and tree embeddings (return as a pair)."""
if self.da_embs:
# DA embeddings
da_repr = self.da_embs.get_embeddings(da)
else:
# DA one-hot representation
da_repr = self.vectorizer.transform([self.da_feats.get_features(tree, {'da': da})])[0]
# tree embeddings
tree_emb_idxs = self.tree_embs.get_embeddings(tree)
return (da_repr, tree_emb_idxs)
def _init_neural_network(self):
# initial layer – tree embeddings & DA 1-hot or embeddings
# input shapes don't contain the batch dimension, but the input Theano types do!
if self.da_embs:
input_shapes = (self.da_embs.get_embeddings_shape(),
self.tree_embs.get_embeddings_shape())
input_types = (T.itensor3, T.itensor3)
layers = [[Embedding('emb_da', self.dict_size, self.emb_size, 'uniform_005'),
Embedding('emb_tree', self.dict_size, self.emb_size, 'uniform_005')]]
else:
input_shapes = ([len(self.vectorizer.get_feature_names())],
self.tree_embs.get_embeddings_shape())
input_types = (T.fmatrix, T.itensor3)
layers = [[Identity('id_da'),
Embedding('emb_tree', self.dict_size, self.emb_size, 'uniform_005')]]
# plain feed-forward networks
if self.nn_shape.startswith('ff'):
layers += [[Flatten('flat_da'), Flatten('flat_tree')], [Concat('concat')]]
num_ff_layers = 2
if self.nn_shape[-1] in ['3', '4']:
num_ff_layers = int(self.nn_shape[-1])
layers += self._ff_layers('ff', num_ff_layers, perc_layer=True)
# convolution with or without max/avg-pooling
elif self.nn_shape.startswith('conv'):
num_conv_layers = 2 if self.nn_shape.startswith('conv2') else 1
pooling = None
if 'maxpool' in self.nn_shape:
pooling = T.max
elif 'avgpool' in self.nn_shape:
pooling = T.mean
if self.da_embs:
da_layers = self._conv_layers('conv_da', num_conv_layers, pooling=pooling)
else:
da_layers = self._id_layers('id_da',
num_conv_layers + (1 if pooling is not None else 0))
tree_layers = self._conv_layers('conv_tree', num_conv_layers, pooling=pooling)
for da_layer, tree_layer in zip(da_layers, tree_layers):
layers.append([da_layer[0], tree_layer[0]])
layers += [[Flatten('flat_da'), Flatten('flat_tree')], [Concat('concat')]]
layers += self._ff_layers('ff', 2, perc_layer=True)
# max-pooling without convolution
elif 'maxpool-ff' in self.nn_shape:
layers += [[Pool1D('mp_da') if self.da_embs else Identity('id_da'),
Pool1D('mp_trees')]
[Concat('concat')], [Flatten('flat')]]
layers += self._ff_layers('ff', 2, perc_layer=True),
# dot-product FF network
elif 'dot' in self.nn_shape:
# with max or average pooling
if 'maxpool' in self.nn_shape or 'avgpool' in self.nn_shape:
pooling = T.mean if 'avgpool' in self.nn_shape else T.max
layers += [[Pool1D('mp_da', pooling_func=pooling)
if self.da_embs else Identity('id_da'),
Pool1D('mp_tree', pooling_func=pooling)]]
layers += [[Flatten('flat_da') if self.da_embs else Identity('id_da'),
Flatten('flat_tree')]]
num_ff_layers = int(self.nn_shape[-1]) if self.nn_shape[-1] in ['2', '3', '4'] else 1
for da_layer, tree_layer in zip(self._ff_layers('ff_da', num_ff_layers),
self._ff_layers('ff_tree', num_ff_layers)):
layers.append([da_layer[0], tree_layer[0]])
layers.append([DotProduct('dot')])
# input: batch * word * sub-embeddings
self.nn = RankNN(layers, input_shapes, input_types, self.normgrad)
log_info("Network shape:\n\n" + str(self.nn))
def _conv_layers(self, name, num_layers=1, pooling=None):
ret = []
for i in xrange(num_layers):
ret.append([Conv1D(name + str(i + 1),
filter_length=self.cnn_filter_length,
num_filters=self.cnn_num_filters,
init=self.init, activation=T.tanh)])
if pooling is not None:
ret.append([Pool1D(name + str(i + 1) + 'pool', pooling_func=pooling)])
return ret
def _id_layers(self, name, num_layers):
ret = []
for i in xrange(num_layers):
ret.append([Identity(name + str(i + 1))])
return ret
def _update_nn(self, bad_feats, good_feats, rate):
"""Changing the NN update call to support arrays of parameters."""
# TODO: this is just adding another dimension to fit the parallelized scoring
# (even if updates are not parallelized). Make it nicer.
bad_feats = ([bad_feats[0]], [bad_feats[1]])
good_feats = ([good_feats[0]], [good_feats[1]])
cost_gcost = self.nn.update(*(bad_feats + good_feats + (rate,)))
log_debug('Cost:' + str(cost_gcost[0]))
param_vals = [param.get_value() for param in self.nn.params]
log_debug('Param norms : ' + str(self._l2s(param_vals)))
log_debug('Gparam norms: ' + str(self._l2s(cost_gcost[1:])))
def _embs_to_str(self):
out = ""
da_emb = self.nn.layers[0][0].e.get_value()
tree_emb = self.nn.layers[0][1].e.get_value()
for idx, emb in enumerate(da_emb):
for key, val in self.dict_slot.items():
if val == idx:
out += key + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for key, val in self.dict_value.items():
if val == idx:
out += key + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for idx, emb in enumerate(tree_emb):
for key, val in self.dict_t_lemma.items():
if val == idx:
out += str(key) + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
for key, val in self.dict_formeme.items():
if val == idx:
out += str(key) + ',' + ','.join([("%f" % d) for d in emb]) + "\n"
return out
def _l2s(self, params):
"""Compute L2-norm of all members of the given list."""
return [np.linalg.norm(param) for param in params]
def store_iter_weights(self):
"""Remember the current weights to be used for averaged perceptron."""
# fh = open('embs.txt', 'a')
# print >> fh, '---', self._embs_to_str()
# fh.close()
self.w_after_iter.append(self.nn.get_param_values())
def score_all(self, trees, da):
cand_embs = [self._extract_feats(tree, da) for tree in trees]
score = self.nn.score([emb[0] for emb in cand_embs], [emb[1] for emb in cand_embs])
return np.atleast_1d(score[0])
class Reranker(object):
def __init__(self, cfg):
self.cfg = cfg
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.mode = cfg.get('mode', 'tokens' if cfg.get('use_tokens') else 'trees')
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.cur_da = None
self.cur_da_bin = None
self.delex_slots = cfg.get('delex_slots', None)
if self.delex_slots:
self.delex_slots = set(self.delex_slots.split(','))
@staticmethod
def get_model_type(cfg):
"""Return the correct model class according to the config."""
if cfg.get('model') == 'e2e_patterns':
from tgen.e2e.slot_error import E2EPatternClassifier
return E2EPatternClassifier
return RerankingClassifier
@staticmethod
def load_from_file(reranker_fname):
"""Detect correct model type and start loading."""
model_type = RerankingClassifier # default to classifier
with file_stream(reranker_fname, 'rb', encoding=None) as fh:
data = pickle.load(fh)
if isinstance(data, type):
from tgen.e2e.slot_error import E2EPatternClassifier
model_type = data
return model_type.load_from_file(reranker_fname)
def save_to_file(self, reranker_fname):
raise NotImplementedError()
def get_all_settings(self):
raise NotImplementedError()
def classify(self, trees):
raise NotImplementedError()
def train(self, das, trees, data_portion=1.0, valid_das=None, valid_trees=None):
raise NotImplementedError()
def _normalize_da(self, da):
if isinstance(da, tuple): # if DA is actually context + DA, ignore context
da = da[1]
if self.delex_slots: # delexicalize the DA if needed
da = da.get_delexicalized(self.delex_slots)
return da
def init_run(self, da):
"""Remember the current DA for subsequent runs of `dist_to_cur_da`."""
self.cur_da = self._normalize_da(da)
da_bin = self.da_vect.transform([self.da_feats.get_features(None, {'da': self.cur_da})])[0]
self.cur_da_bin = da_bin != 0
def dist_to_da(self, da, trees, return_classif=False):
"""Return Hamming distance of given trees to the given DA.
@param da: the DA as the base of the Hamming distance measure
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree (+ resulting classification if return_classif)
"""
self.init_run(da)
ret = self.dist_to_cur_da(trees, return_classif)
self.cur_da = None
self.cur_da_bin = None
return ret
def dist_to_cur_da(self, trees, return_classif=False):
"""Return Hamming distance of given trees to the current DA (set in `init_run`).
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree (+ resulting classification if return_classif)
"""
da_bin = self.cur_da_bin
covered = self.classify(trees)
dist = [sum(abs(c - da_bin)) for c in covered]
if return_classif:
return dist, [[f for f, c_ in zip(self.da_vect.feature_names_, c) if c_] for c in covered]
return dist
class RerankingClassifier(TFModel):
"""A classifier for trees that decides which DAIs are currently represented
(to be used in limiting candidate generator and/or re-scoring the trees)."""
def __init__(self, cfg):
super(RerankingClassifier, self).__init__(scope_name='rerank-' +
cfg.get('scope_suffix', ''))
self.cfg = cfg
self.language = cfg.get('language', 'en')
self.selector = cfg.get('selector', '')
self.tree_embs = cfg.get('nn', '').startswith('emb')
if self.tree_embs:
self.tree_embs = TreeEmbeddingClassifExtract(cfg)
self.emb_size = cfg.get('emb_size', 50)
self.mode = cfg.get('mode', 'tokens' if cfg.get('use_tokens') else 'trees')
self.nn_shape = cfg.get('nn_shape', 'ff')
self.num_hidden_units = cfg.get('num_hidden_units', 512)
self.passes = cfg.get('passes', 200)
self.min_passes = cfg.get('min_passes', 0)
self.alpha = cfg.get('alpha', 0.1)
self.randomize = cfg.get('randomize', True)
self.batch_size = cfg.get('batch_size', 1)
self.validation_freq = cfg.get('validation_freq', 10)
self.max_cores = cfg.get('max_cores')
self.cur_da = None
self.cur_da_bin = None
self.checkpoint_path = None
self.delex_slots = cfg.get('delex_slots', None)
if self.delex_slots:
self.delex_slots = set(self.delex_slots.split(','))
# Train Summaries
self.train_summary_dir = cfg.get('tb_summary_dir', None)
if self.train_summary_dir:
self.loss_summary_reranker = None
self.train_summary_op = None
self.train_summary_writer = None
def save_to_file(self, model_fname):
"""Save the classifier to a file (actually two files, one for configuration and one
for the TensorFlow graph, which must be stored separately).
@param model_fname: file name (for the configuration file); TF graph will be stored with a \
different extension
"""
model_fname = self.tf_check_filename(model_fname)
log_info("Saving classifier to %s..." % model_fname)
with file_stream(model_fname, 'wb', encoding=None) as fh:
pickle.dump(self.get_all_settings(), fh, protocol=pickle.HIGHEST_PROTOCOL)
tf_session_fname = re.sub(r'(.pickle)?(.gz)?$', '.tfsess', model_fname)
if hasattr(self, 'checkpoint_path') and self.checkpoint_path:
self.restore_checkpoint()
shutil.rmtree(os.path.dirname(self.checkpoint_path))
self.saver.save(self.session, tf_session_fname)
def get_all_settings(self):
"""Get all settings except the trained model parameters (to be stored in a pickle)."""
data = {'cfg': self.cfg,
'da_feats': self.da_feats,
'da_vect': self.da_vect,
'tree_embs': self.tree_embs,
'input_shape': self.input_shape,
'num_outputs': self.num_outputs, }
if self.tree_embs:
data['dict_size'] = self.dict_size
else:
data['tree_feats'] = self.tree_feats
data['tree_vect'] = self.tree_vect
return data
def _save_checkpoint(self):
"""Save a checkpoint to a temporary path; set `self.checkpoint_path` to the path
where it is saved; if called repeatedly, will always overwrite the last checkpoint."""
if not self.checkpoint_path:
path = tempfile.mkdtemp(suffix="", prefix="tftreecl-")
self.checkpoint_path = os.path.join(path, "ckpt")
log_info('Saving checkpoint to %s' % self.checkpoint_path)
self.saver.save(self.session, self.checkpoint_path)
def restore_checkpoint(self):
if not self.checkpoint_path:
return
self.saver.restore(self.session, self.checkpoint_path)
@staticmethod
def load_from_file(model_fname):
"""Load the reranker from a file (actually two files, one for configuration and one
for the TensorFlow graph, which must be stored separately).
@param model_fname: file name (for the configuration file); TF graph must be stored with a \
different extension
"""
log_info("Loading reranker from %s..." % model_fname)
with file_stream(model_fname, 'rb', encoding=None) as fh:
data = pickle.load(fh)
ret = RerankingClassifier(cfg=data['cfg'])
ret.load_all_settings(data)
# re-build TF graph and restore the TF session
tf_session_fname = os.path.abspath(re.sub(r'(.pickle)?(.gz)?$', '.tfsess', model_fname))
ret._init_neural_network()
ret.saver.restore(ret.session, tf_session_fname)
return ret
def train(self, das, trees, data_portion=1.0, valid_das=None, valid_trees=None):
"""Run training on the given training data.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (defaults to 1.0)
@param valid_das: validation data DAs
@param valid_trees: list of lists of corresponding paraphrases (same length as valid_das)
"""
log_info('Training reranking classifier...')
# initialize training
self._init_training(das, trees, data_portion)
if self.mode in ['tokens', 'tagged_lemmas'] and valid_trees is not None:
valid_trees = [self._tokens_to_flat_trees(paraphrases,
use_tags=self.mode == 'tagged_lemmas')
for paraphrases in valid_trees]
# start training
top_comb_cost = float('nan')
for iter_no in xrange(1, self.passes + 1):
self.train_order = range(len(self.train_trees))
if self.randomize:
rnd.shuffle(self.train_order)
pass_cost, pass_diff = self._training_pass(iter_no)
if self.validation_freq and iter_no > self.min_passes and iter_no % self.validation_freq == 0:
valid_diff = 0
if valid_das:
valid_diff = np.sum([np.sum(self.dist_to_da(d, t))
for d, t in zip(valid_das, valid_trees)])
# cost combining validation and training data performance
# (+ "real" cost with negligible weight)
comb_cost = 1000 * valid_diff + 100 * pass_diff + pass_cost
log_info('Combined validation cost: %8.3f' % comb_cost)
# if we have the best model so far, save it as a checkpoint (overwrite previous)
if math.isnan(top_comb_cost) or comb_cost < top_comb_cost:
top_comb_cost = comb_cost
self._save_checkpoint()
# restore last checkpoint (best performance on devel data)
self.restore_checkpoint()
def classify(self, trees):
"""Classify the tree -- get DA slot-value pairs and DA type to which the tree
corresponds (as 1/0 array).
"""
if self.tree_embs:
inputs = np.array([self.tree_embs.get_embeddings(tree) for tree in trees])
else:
inputs = self.tree_vect.transform([self.tree_feats.get_features(tree, {})
for tree in trees])
fd = {}
self._add_inputs_to_feed_dict(inputs, fd)
results = self.session.run(self.outputs, feed_dict=fd)
# normalize & binarize the result
return np.array([[1. if r > 0 else 0. for r in result] for result in results])
def _normalize_da(self, da):
if isinstance(da, tuple): # if DA is actually context + DA, ignore context
da = da[1]
if self.delex_slots: # delexicalize the DA if needed
da = da.get_delexicalized(self.delex_slots)
return da
def init_run(self, da):
"""Remember the current DA for subsequent runs of `dist_to_cur_da`."""
self.cur_da = self._normalize_da(da)
da_bin = self.da_vect.transform([self.da_feats.get_features(None, {'da': self.cur_da})])[0]
self.cur_da_bin = da_bin != 0
def dist_to_da(self, da, trees):
"""Return Hamming distance of given trees to the given DA.
@param da: the DA as the base of the Hamming distance measure
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree
"""
da = self._normalize_da(da)
da_bin = self.da_vect.transform([self.da_feats.get_features(None, {'da': da})])[0]
da_bin = da_bin != 0
covered = self.classify(trees)
return [sum(abs(c - da_bin)) for c in covered]
def dist_to_cur_da(self, trees):
"""Return Hamming distance of given trees to the current DA (set in `init_run`).
@param trees: list of trees to measure the distance
@return: list of Hamming distances for each tree
"""
da_bin = self.cur_da_bin
covered = self.classify(trees)
return [sum(abs(c - da_bin)) for c in covered]
def _init_training(self, das, trees, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (0.0-1.0)
"""
# read input from files or take it directly from parameters
if not isinstance(das, list):
log_info('Reading DAs from ' + das + '...')
das = read_das(das)
if not isinstance(trees, list):
log_info('Reading t-trees from ' + trees + '...')
ttree_doc = read_ttrees(trees)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tls, use_tags=True)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
elif self.mode in ['tokens', 'tagged_lemmas']:
trees = self._tokens_to_flat_trees(trees, use_tags=self.mode == 'tagged_lemmas')
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# ignore contexts, if they are contained in the DAs
if isinstance(self.train_das[0], tuple):
self.train_das = [da for (context, da) in self.train_das]
# delexicalize if DAs are lexicalized and we don't want that
if self.delex_slots:
self.train_das = [da.get_delexicalized(self.delex_slots) for da in self.train_das]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([self.tree_embs.get_embeddings(tree) for tree in self.train_trees])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.X = [self.tree_feats.get_features(tree, {}) for tree in self.train_trees]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [self.da_feats.get_features(None, {'da': da}) for da in self.train_das]
self.y = self.da_vect.fit_transform(self.y)
log_info('Number of binary classes: %d.' % len(self.da_vect.get_feature_names()))
# initialize I/O shapes
if not self.tree_embs:
self.input_shape = list(self.X[0].shape)
else:
self.input_shape = self.tree_embs.get_embeddings_shape()
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
# initialize the NN variables
self.session.run(tf.global_variables_initializer())
def _tokens_to_flat_trees(self, sents, use_tags=False):
"""Use sentences (pairs token-tag) read from Treex files and convert them into flat
trees (each token has a node right under the root, lemma is the token, formeme is 'x').
Uses TokenEmbeddingSeq2SeqExtract conversion there and back.
@param sents: sentences to be converted
@param use_tags: use tags in the embeddings? (only for lemma-tag pairs in training, \
not testing)
@return: a list of flat trees
"""
tree_embs = (TokenEmbeddingSeq2SeqExtract(cfg=self.cfg)
if not use_tags
else TaggedLemmasEmbeddingSeq2SeqExtract(cfg=self.cfg))
tree_embs.init_dict(sents)
# no postprocessing, i.e. keep lowercasing/plural splitting if set in the configuration
return [tree_embs.ids_to_tree(tree_embs.get_embeddings(sent), postprocess=False)
for sent in sents]
def _init_neural_network(self):
"""Create the neural network for classification, according to the self.nn_shape
parameter (as set in configuration)."""
# set TensorFlow random seed
tf.set_random_seed(rnd.randint(-sys.maxint, sys.maxint))
self.targets = tf.placeholder(tf.float32, [None, self.num_outputs], name='targets')
with tf.variable_scope(self.scope_name):
# feedforward networks
if self.nn_shape.startswith('ff'):
self.inputs = tf.placeholder(tf.float32, [None] + self.input_shape, name='inputs')
num_ff_layers = 2
if self.nn_shape[-1] in ['0', '1', '3', '4']:
num_ff_layers = int(self.nn_shape[-1])
self.outputs = self._ff_layers('ff', num_ff_layers, self.inputs)
# RNNs
elif self.nn_shape.startswith('rnn'):
self.initial_state = tf.placeholder(tf.float32, [None, self.emb_size])
self.inputs = [tf.placeholder(tf.int32, [None], name=('enc_inp-%d' % i))
for i in xrange(self.input_shape[0])]
self.cell = tf.contrib.rnn.BasicLSTMCell(self.emb_size)
self.outputs = self._rnn('rnn', self.inputs)
# the cost as computed by TF actually adds a "fake" sigmoid layer on top
# (or is computed as if there were a sigmoid layer on top)
self.cost = tf.reduce_mean(tf.reduce_sum(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.outputs, labels=self.targets, name='CE'), 1))
# NB: this would have been the "true" cost function, if there were a "real" sigmoid layer on top.
# However, it is not numerically stable in practice, so we have to use the TF function.
# self.cost = tf.reduce_mean(tf.reduce_sum(self.targets * -tf.log(self.outputs)
# + (1 - self.targets) * -tf.log(1 - self.outputs), 1))
self.optimizer = tf.train.AdamOptimizer(self.alpha)
self.train_func = self.optimizer.minimize(self.cost)
# Tensorboard summaries
if self.train_summary_dir:
self.loss_summary_reranker = tf.summary.scalar("loss_reranker", self.cost)
self.train_summary_op = tf.summary.merge([self.loss_summary_reranker])
# initialize session
session_config = None
if self.max_cores:
session_config = tf.ConfigProto(inter_op_parallelism_threads=self.max_cores,
intra_op_parallelism_threads=self.max_cores)
self.session = tf.Session(config=session_config)
# this helps us load/save the model
self.saver = tf.train.Saver(tf.global_variables())
if self.train_summary_dir: # Tensorboard summary writer
self.train_summary_writer = tf.summary.FileWriter(
os.path.join(self.train_summary_dir, "reranker"), self.session.graph)
def _ff_layers(self, name, num_layers, X):
width = [np.prod(self.input_shape)] + (num_layers * [self.num_hidden_units]) + [self.num_outputs]
# the last layer should be a sigmoid, but TF simulates it for us in cost computation
# so the output is "unnormalized sigmoids"
activ = (num_layers * [tf.tanh]) + [tf.identity]
Y = X
for i in xrange(num_layers + 1):
w = tf.get_variable(name + ('-w%d' % i), (width[i], width[i + 1]),
initializer=tf.random_normal_initializer(stddev=0.1))
b = tf.get_variable(name + ('-b%d' % i), (width[i + 1],),
initializer=tf.constant_initializer())
Y = activ[i](tf.matmul(Y, w) + b)
return Y
def _rnn(self, name, enc_inputs):
encoder_cell = tf.contrib.rnn.EmbeddingWrapper(self.cell, self.dict_size, self.emb_size)
encoder_outputs, encoder_state = tf.contrib.rnn.static_rnn(encoder_cell, enc_inputs, dtype=tf.float32)
# TODO for historical reasons, the last layer uses both output and state.
# try this just with outputs (might work exactly the same)
if isinstance(self.cell.state_size, tf.contrib.rnn.LSTMStateTuple):
state_size = self.cell.state_size.c + self.cell.state_size.h
final_input = tf.concat(axis=1, values=encoder_state) # concat c + h
else:
state_size = self.cell.state_size
final_input = encoder_state
w = tf.get_variable(name + '-w', (state_size, self.num_outputs),
initializer=tf.random_normal_initializer(stddev=0.1))
b = tf.get_variable(name + 'b', (self.num_outputs,), initializer=tf.constant_initializer())
return tf.matmul(final_input, w) + b
def _batches(self):
"""Create batches from the input; use as iterator."""
for i in xrange(0, len(self.train_order), self.batch_size):
yield self.train_order[i: i + self.batch_size]
def _add_inputs_to_feed_dict(self, inputs, fd):
if self.nn_shape.startswith('rnn'):
fd[self.initial_state] = np.zeros([inputs.shape[0], self.emb_size])
sliced_inputs = np.squeeze(np.array(np.split(np.array([ex for ex in inputs
if ex is not None]),
len(inputs[0]), axis=1)), axis=2)
for input_, slice_ in zip(self.inputs, sliced_inputs):
fd[input_] = slice_
else:
fd[self.inputs] = inputs
def _training_pass(self, pass_no):
"""Perform one training pass through the whole training data, print statistics."""
pass_start_time = time.time()
log_debug('\n***\nTR %05d:' % pass_no)
log_debug("Train order: " + str(self.train_order))
pass_cost = 0
pass_diff = 0
for tree_nos in self._batches():
log_debug('TREE-NOS: ' + str(tree_nos))
log_debug("\n".join(unicode(self.train_trees[i]) + "\n" + unicode(self.train_das[i])
for i in tree_nos))
log_debug('Y: ' + str(self.y[tree_nos]))
fd = {self.targets: self.y[tree_nos]}
self._add_inputs_to_feed_dict(self.X[tree_nos], fd)
if self.train_summary_dir: # also compute Tensorboard summaries
results, cost, _, train_summary_op = self.session.run(
[self.outputs, self.cost, self.train_func, self.train_summary_op], feed_dict=fd)
else:
results, cost, _ = self.session.run([self.outputs, self.cost, self.train_func],
feed_dict=fd)
bin_result = np.array([[1. if r > 0 else 0. for r in result] for result in results])
log_debug('R: ' + str(bin_result))
log_debug('COST: %f' % cost)
log_debug('DIFF: %d' % np.sum(np.abs(self.y[tree_nos] - bin_result)))
pass_cost += cost
pass_diff += np.sum(np.abs(self.y[tree_nos] - bin_result))
# print and return statistics
self._print_pass_stats(pass_no, datetime.timedelta(seconds=(time.time() - pass_start_time)),
pass_cost, pass_diff)
if self.train_summary_dir: # Tensorboard: iteration summary
self.train_summary_writer.add_summary(train_summary_op, pass_no)
return pass_cost, pass_diff
def _print_pass_stats(self, pass_no, time, cost, diff):
log_info('PASS %03d: duration %s, cost %f, diff %d' % (pass_no, str(time), cost, diff))
def evaluate_file(self, das_file, ttree_file):
"""Evaluate the reranking classifier on a given pair of DA/tree files (show the
total Hamming distance and total number of DAIs)
@param das_file: DA file path
@param ttree_file: trees/sentences file path
@return: a tuple (total DAIs, distance)
"""
das = read_das(das_file)
ttree_doc = read_ttrees(ttree_file)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tls)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
da_len = 0
dist = 0
for da, tree in zip(das, trees):
da_len += len(da)
dist += self.dist_to_da(da, [tree])[0]
return da_len, dist
def _init_training(self, das, trees, data_portion):
"""Initialize training.
Store input data, initialize 1-hot feature representations for input and output and
transform training data accordingly, initialize the classification neural network.
@param das: name of source file with training DAs, or list of DAs
@param trees: name of source file with corresponding trees/sentences, or list of trees
@param data_portion: portion of the training data to be used (0.0-1.0)
"""
# read input from files or take it directly from parameters
if not isinstance(das, list):
log_info('Reading DAs from ' + das + '...')
das = read_das(das)
if not isinstance(trees, list):
log_info('Reading t-trees from ' + trees + '...')
ttree_doc = read_ttrees(trees)
if self.mode == 'tokens':
tokens = tokens_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tokens)
elif self.mode == 'tagged_lemmas':
tls = tagged_lemmas_from_doc(ttree_doc, self.language, self.selector)
trees = self._tokens_to_flat_trees(tls, use_tags=True)
else:
trees = trees_from_doc(ttree_doc, self.language, self.selector)
elif self.mode in ['tokens', 'tagged_lemmas']:
trees = self._tokens_to_flat_trees(trees, use_tags=self.mode == 'tagged_lemmas')
# make training data smaller if necessary
train_size = int(round(data_portion * len(trees)))
self.train_trees = trees[:train_size]
self.train_das = das[:train_size]
# ignore contexts, if they are contained in the DAs
if isinstance(self.train_das[0], tuple):
self.train_das = [da for (context, da) in self.train_das]
# delexicalize if DAs are lexicalized and we don't want that
if self.delex_slots:
self.train_das = [da.get_delexicalized(self.delex_slots) for da in self.train_das]
# add empty tree + empty DA to training data
# (i.e. forbid the network to keep any of its outputs "always-on")
train_size += 1
self.train_trees.append(TreeData())
empty_da = DA.parse('inform()')
self.train_das.append(empty_da)
self.train_order = range(len(self.train_trees))
log_info('Using %d training instances.' % train_size)
# initialize input features/embeddings
if self.tree_embs:
self.dict_size = self.tree_embs.init_dict(self.train_trees)
self.X = np.array([self.tree_embs.get_embeddings(tree) for tree in self.train_trees])
else:
self.tree_feats = Features(['node: presence t_lemma formeme'])
self.tree_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.X = [self.tree_feats.get_features(tree, {}) for tree in self.train_trees]
self.X = self.tree_vect.fit_transform(self.X)
# initialize output features
self.da_feats = Features(['dat: dat_presence', 'svp: svp_presence'])
self.da_vect = DictVectorizer(sparse=False, binarize_numeric=True)
self.y = [self.da_feats.get_features(None, {'da': da}) for da in self.train_das]
self.y = self.da_vect.fit_transform(self.y)
log_info('Number of binary classes: %d.' % len(self.da_vect.get_feature_names()))
# initialize I/O shapes
if not self.tree_embs:
self.input_shape = list(self.X[0].shape)
else:
self.input_shape = self.tree_embs.get_embeddings_shape()
self.num_outputs = len(self.da_vect.get_feature_names())
# initialize NN classifier
self._init_neural_network()
# initialize the NN variables
self.session.run(tf.global_variables_initializer())